added reference to canary-flash paper

README.md

@@ -214,7 +214,7 @@ NVIDIA NeMo Canary [1] is a family of multilingual multi-tasking models that ach
 
 
 ## Model Architecture:
-Canary is an encoder-decoder model with FastConformer [2] Encoder and Transformer Decoder [3]. With audio features extracted from the encoder, task tokens such as \<target language\>, \<task\>, \<toggle timestamps\> and \<toggle PnC\> are fed into the Transformer Decoder to trigger the text generation process. Canary uses a concatenated tokenizer [4] from individual SentencePiece [5] tokenizers of each language, which makes it easy to scale up to more languages. The canary-1b-flash model has 32 encoder layers and 4 layers of decoder layers, leading to a total of 883M parameters.
+Canary is an encoder-decoder model with FastConformer [2] Encoder and Transformer Decoder [3]. With audio features extracted from the encoder, task tokens such as \<target language\>, \<task\>, \<toggle timestamps\> and \<toggle PnC\> are fed into the Transformer Decoder to trigger the text generation process. Canary uses a concatenated tokenizer [4] from individual SentencePiece [5] tokenizers of each language, which makes it easy to scale up to more languages. The canary-1b-flash model has 32 encoder layers and 4 layers of decoder layers, leading to a total of 883M parameters. For more details about the architecture, please refer to [9].
 
 ## NVIDIA NeMo
 
@@ -399,7 +399,7 @@ Model Fairness:
 
 ## Training
 
-canary-1b-flash is trained using the NVIDIA NeMo toolkit [6] for a total of 200K steps with 2D bucketing [7] and optimal batch sizes set using OOMptimizer [7].The model is trained on 128 NVIDIA A100 80GB GPUs. 
+canary-1b-flash is trained using the NVIDIA NeMo toolkit [6] for a total of 200K steps with 2D bucketing [9] and optimal batch sizes set using OOMptimizer [7].The model is trained on 128 NVIDIA A100 80GB GPUs. 
 The model can be trained using this [example script](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/speech_multitask/speech_to_text_aed.py) and [base config](https://github.com/NVIDIA/NeMo/blob/main/examples/asr/conf/speech_multitask/fast-conformer_aed.yaml).
 
 The tokenizers for these models were built using the text transcripts of the train set with this [script](https://github.com/NVIDIA/NeMo/blob/main/scripts/tokenizers/process_asr_text_tokenizer.py).
@@ -425,7 +425,6 @@ WER on [HuggingFace OpenASR leaderboard](https://huggingface.co/spaces/hf-audio/
 |:---------:|:-----------:|:------:|:------:|:------:|:------:|:------:|:------:|:------:|:------:|:------:|
 | 2.2.0  | canary-1b-flash | 928.19 | 13.08 | 9.88 | 1.48 | 2.87 | 12.77 | 1.95 | 3.09 | 5.64 |
 
-
 WER on [MLS](https://huggingface.co/datasets/facebook/multilingual_librispeech) test set:
 
 | **Version** | **Model**  | **De**   | **Es**   | **Fr**   |
@@ -516,6 +515,8 @@ canary-1b-flash is released under the CC-BY-4.0 license. By using this model, yo
 
 [8] [Towards Measuring Fairness in AI: the Casual Conversations Dataset](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9634168)
 
+[9] [Training and Inference Efficiency of Encoder-Decoder Speech Models](https://arxiv.org/pdf/2503.05931)
+
 ## Ethical Considerations:
 NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications.  When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.  
 Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).